Predicting wealth score from remote sensing satellite images and household survey data using deep learning

The most exigent call of the United Nations’ 17 sustainable goals is to end poverty every-where by 2030. Unlike in the past, when poverty was measured based on data collectedthrough ground-level surveys, the new technology adopted by many developing and devel-oped countries is to estimate the poverty index using remote sensing satellite images withthe help of machine learning techniques. Our approach demonstrates the prediction ofcluster wealth score and establishes the relationship between wealth score obtained fromDemographic and Health Survey (DHS) data and remote sensing satellite images of In-dia by calculating Pearson’s correlation coefficient (r2). The implementation results havebeen analyzed in four phases. Phase 1 comprises four regression models (RMs): Ridge,RANSAC, Lasso, andk-nearest neighbor for feature extraction from a pre-trained con-volutional neural network model using daylight & nightlight images. Here, the Lasso RMoutperforms the others and is best suited for predicting the wealth score. Phase 2 cat-egorizes daylight images with DHS data, where the Lasso RM efficiently generates thecluster wealth score. Phase 3 focuses on images of specific regions of Delhi, Tamil Nadu, Maharashtra and Telangana, using the Lasso RM, as it emerged as the best predictor ofcluster wealth score in the previous two phases. Phase 4 compares the results attainedthrough our proposed model with existing results.